Across regulatory change, risk, finance, governance and delivery, the work today is increasingly narrative-driven, interpretation-heavy, overwhelmingly text-based, and documented for audit rather than speed. A large part of the job is no longer executing transactions, but reading, interpreting, aligning and documenting.
Without explicitly deciding to, banks have become knowledge factories, not transaction factories.
This is why the current wave of GenAI is often misunderstood. It is not revolutionary because it suddenly “thinks like a human”. It is impactful because the work itself has evolved into something GenAI is inherently good at.
Modern banking work requires processing large volumes of text, interpreting meaning consistently across documents, drafting and redrafting content, and retaining context across iterations and stakeholders. These are exactly the capabilities GenAI brings.
At the same time, banks struggle to scale one thing: expertise. Senior experts spend a disproportionate amount of time rereading similar material, re-explaining decisions, rewriting comparable content, and revalidating what was already agreed. The bottleneck is not decision-making. It is getting to shared understanding.
GenAI does not replace judgement. It collapses the cost of understanding. And that is why this moment matters. Not because AI has fundamentally changed, but because banking work has quietly become structured in a way that allows AI to plug in directly.
Despite the clarity of the opportunity, the majority of financial institutions are still early-stage when it comes to operational adoption. And when you look closely at why, three patterns keep appearing.
The first is waiting. The instinct to hold back until conditions are perfect is understandable in a regulated environment. Regulatory frameworks are still evolving. The technology keeps moving. The stakes are high. But waiting does not reduce risk. It relocates it. In practice, it pushes GenAI activity into the shadow: employees using personal devices outside any oversight or organisational learning. The institution does not stop experimenting. It just loses visibility over what is being done.
The second is pilot fragmentation. Many institutions have run multiple proofs of concept, each with a sponsor, a use case, and an impressive demo. But somewhere between proof of concept and production, progress stalls. Not because the technology failed. Not because the use case was wrong. But because nobody defined what came next. No operating model, no named owner for adoption, no shared measure of success. The pilot was designed to prove the technology. Not to survive the organisation.
The third is tool-led thinking. With Gen-AI in particular, there is a strong pull to resolve platform, partnership, and data architecture questions before the business has generated enough evidence of value, or has defined their target vision/requirements. The instinct is understandable: IT teams are accountable for resilience, third-party risk under DORA, and the eventual cost of scale, and a coherent foundation feels like the responsible starting point. But pre-committing to a shared data layer, a primary model partner, or an enterprise-wide orchestration platform before pilots have run forces strategic decisions in an information vacuum. Business teams wait for foundations they cannot yet specify; technology teams build for requirements that have not yet surfaced. What we commonly see are well-intentioned programmes attempting to answer enterprise-architecture questions during what should be cheap, fast learning cycles. The result is slower pilots, higher sunk cost, and platform choices that often need to be revisited once the actual use cases reveal themselves.
The organisations making genuine progress share one thing: they did not try to solve the technology challenge, the governance challenge, and the people challenge one at a time. They moved across all of these dimensions in parallel.
Working with financial institutions on exactly this challenge, we have found that six weeks is enough time to move from a broad GenAI ambition to a prioritised, defensible portfolio of use cases ready to mobilise. Not a roadmap in theory. A portfolio in practice, with a clear rationale for what to pursue first, what to defer, and what to set aside entirely.
The key is a process that rates each use case transparently against measurable criteria: strategic relevance, feasibility, risk proportionality, and return on investment. In a regulated environment, the decision about which use cases to move forward with needs to be explainable, not just to leadership but to the second line and to auditors. Defensibility is not a compliance requirement bolted on afterwards. It is a design principle from the start.
We structure this as a dual-velocity approach. One track starts fast: a small number of internal, low-risk use cases that deliver visible productivity gains and give senior experts hands-on exposure to what GenAI actually does in practice. The other builds the foundations in parallel: governance frameworks, data controls, reusable architecture patterns, and clear lines of accountability. Not one, then the other. Both, at once.
The goal is not transformation at scale from day one. It is controlled learning that compounds. Because what we consistently find is that the first use case going live is rarely the thing that unlocks momentum. What unlocks momentum is the organisation finally having a shared language, a transparent basis for its decisions, and a sequenced plan that leadership can stand behind.
Humans remain accountable, decision-makers and supervisors throughout. But they no longer need to carry the full burden of getting to understanding.
If the ambition is clear but the path is not, that is exactly where a structured, deliberate start delivers the most value. Start small, experiment deliberately, and build an approach that is controlled, defensible, and built to scale.